A Novel Approach For Knowledge Discovery From Multiple Large Tables Of Relational Database Systems

Abstract

In this thesis, a new framework is proposed for extracting knowledge from data stored in multiple normalized relational tables connected to each other by key attributes. Previous methods for knowledge discovery in database were based on single flat tables [Han and Fu 96, Srikant and Agrawal 96a, Mehta et al. 96, Chen et al.•96, Fayyad 97] and these methods cannot be used for extracting patterns that reflect the relationship between structured properties. So a lot of useful patterns cannot be extracted using these methods. The representation language in the proposed framework treats the data as objects and associated properties of nested structure that describe them. The extracted knowledge shows the relationship among certain properties about a subset of objects represented in the database. The definition of association rules is extended with a new certainty measure called backward confidence. Also several disjuncts are allowed in the extended rules that are called mutual association rules. These extensions were not defined in the previous definition of association rules. Using the extended definition, several types of extracted knowledge can be mapped to mutual association rules. Using the proposed framework two new schemes have been developed and •implemented. The first scheme is used for extracting any type of frequent patterns that describe the relationship among several types of properties about the underlined objects stored in the database. The discovery task is directed by a rule template query that restricts the search for only the required patterns. The extracted frequent patterns have a restricted form of mutual association rules. Several algorithms for extracting frequent patterns according to the user query have been developed. The second scheme is used for extracting characteristic descriptions about a subset of objects represented in the database.